When setting up spectral windows in the Observing Tool, users have the option to use Spectral Averaging by factors of 2, 4, 8 and 16. This has the advantage of reducing the data rate (see "How can I reduce my data rates for an observation (and reduce the size of my dataset)?"), and can also improve the ease of working with and analyzing the final imaging products. The highest spectral resolution correlator mode -- Frequency Division Mode (FDM) -- will result in data with 3840 channels when no averaging is selected. The resulting image cubes can be as large as tens of GB, which can be very cumbersome to work with. For example, using a machine with 64 GB RAM, a 34 GB cube takes about 3.5 minutes to load in the CASA viewer. If the number of channels were reduced by the use of spectral averaging, all image-based operations (including cleaning, viewing, etc) will be considerably shortened.
The primary consideration for selecting spectral averaging is, what spectral resolution is needed for your science questions? You should try to get 3-5 spectral resolution elements across the line-widths of interest (although detection experiments can afford poorer resolutions, but not poorer than the expected line-width or the signal-to-noise will be reduced). Note that for ALMA by default all observations are Hanning smoothed, meaning that spectral averaging by a factor of 2 can be used with only a very modest reduction in the spectral resolution. See section 5.5.2 Spectral Setup of the Cycle 3 Technical Handbook for more information.
Spectral averaging can either be specified during Phase 1 (proposal submission), but also at Phase 2 (Scheduling Block generation). However, once selected it cannot be reversed.
A note on Spectral Response provides a deeper look at channel averaging and the Spectral Response Function.